Research map / Architectures and efficiency / Mixture of experts
TA-MoE: Topology-Aware Large Scale Mixture-of-Expert Training
Chang Chen, Min Li, Zhihua Wu, Dianhai Yu, Chao Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| uneven dispatch (more data to neighbor device) instead of even dispatch | MoE all-to-all · [2,2] symmetric tree, 128MB | communication time (us): 14019 -> 10765 | Improved |
| topology-aware loss instead of load-balance loss | FastMoE · openwebtext2 8 experts | valid PPL: 18.12 -> 17.97 | Improved |
| topology-aware loss instead of load-balance loss | FastMoE · openwebtext2 16 experts | valid PPL: 15.39 -> 15.18 | Improved |
| topology-aware loss instead of load-balance loss | FastMoE · openwebtext2 32 experts | valid PPL: 13.53 -> 13.37 | Improved |
| topology-aware loss instead of load-balance loss | FastMoE · openwebtext2 48 experts | valid PPL: 12.49 -> 12.55 | Hurt |
| add TA-MoE to DeepSpeed-MoE | DeepSpeed-MoE · GPT MoE across clusters A/B/C | throughput speedup: 1.05x-1.61x | Improved |
| add TA-MoE to FastMoE | FastMoE · GPT MoE across clusters A/B/C | throughput speedup: 1.01x-4.77x | Improved |
| TA-MoE on cluster C with cross-switch contention | FastMoE · cluster C | throughput speedup: 4.77x | Improved |
| TA-MoE adaptive loss instead of FasterMoE Hir compulsory dispatch | FasterMoE · cluster C, val loss 3.1 | time to convergence: 1.25x | Improved |
| TA-MoE adaptive loss instead of FasterMoE Hir compulsory dispatch | FasterMoE · cluster C, val loss 2.9 | time to convergence: 1.47x | Improved |
| TA-MoE adaptive loss instead of FasterMoE Hir compulsory dispatch | FasterMoE · cluster C, val loss 2.8 | time to convergence: 1.54x | Improved |
| TA-MoE on communication cost | FastMoE · cluster C | communication cost speedup: 1.16x to 6.4x | Improved |
Closest papers by meaning
- Dense Backpropagation Improves Training for Sparse Mixture-of-Experts (NeurIPS 2025 · similarity 0.88)
- LSH-MoE: Communication-efficient MoE Training via Locality-Sensitive Hashing (NeurIPS 2024 · similarity 0.88)
- Scaling Beyond the GPU Memory Limit for Large Mixture-of-Experts Model Training (ICML 2024 · similarity 0.87)
- MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts (ICLR 2025 · similarity 0.87)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 2025 · similarity 0.86)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
